I would suggest to use a custom CDF file instead. This will solve all your questions.

In a few words a couple of ressources propose to use some clean annotation for a lot of microarrays technologies and versions. One of the most used is the BrainArray CDF files (http://brainarray.mbni.med.umich.edu/Brainarray/Database/CustomCDF/genomic_curated_CDF.asp)

or for you

http://brainarray.mbni.med.umich.edu/Brainarray/Database/CustomCDF/14.1.0/entrezg.asp

At the end, for a given gene, all probes (inside a probe-set) were tested, and validated to propose a new (custom) annotation per gene.

This can also easily be used through R and BioConductor

Never average probe sets for the same gene- they are designed to report on different regions of the mRNA which may correspond to alternate transcripts or splice forms, and will not necessarily correlate. Always treat each probe set as an independent metric.

Thanks Dr. Daniel for answering the second query of mine. Do you have any idea about the different type of probe sets that I have mentioned above in my question section and which type of probe sets to consider. My specific problem is a gene with alternate transcripts have been assigned with different ENST and ENSP IDs (ensemble). If I am getting two probe sets which represents the same ENSP/ ENST IDs (same isoforms vs two probe sets) and each probe sets having different expression values (one is up-regulated and the other one is no change in expression) then how to consider these probe sets values to represent the transcript expression. Hope I made the query very clear to you.

Hi Amit. It is not unusual to observe discordant results for the same target gene with different probe sets. Microarrays are simply screening tools and yield false positives and false negatives for a host of different reasons (e.g. target sequence cross-hybridizes to a different gene, annotations in genomes change or are in error, spots are misread, etc). If you think your gene is biologically interesting, check its expression by real-time PCR. That is the only definitive way to know if the indication of a change from your microarray dataset is real or artifact.

In case of multiple probe sets mapping to the same transcript one might consider taking only the prob set with the highest variance across the arrays. More information on this kind of filtering can be found in the bioconductor package 'genefilter' function nsFilter.

As regards your first question: here is an explanation for the nomenclature http://www.affymetrix.com/support/help/faqs/mouse_430/faq_8.jsp

In practice, I would consider all probe sets in an initial analysis,

Another note of caution interpreting your data is that hybridization/signal strength can affect the sensitivity of your read-out of fold change. Very weak probe sets can fail to show induction of gene expression, and very strong ones can saturate and thereby under report gene activation. If you select for highest variance, you may also be inadvertently introducing a selection bias for signal strength (more likely to see downregulated probe sets with mid-high signal strength, and upregulated at low-to-moderate signal strength).

Therefore, I second Tobias' last point. Stick with all probe sets in your initial analyses.

Thanks Dr. Daniel and Dr. Straub, For your kind Information I am trying to map the expression levels of probe sets to human whole genome network. This is where I am getting the problem. Although I am trying to consider all the prob sets representing a single genes where many are showing the same pattern of expression (Tight correlation) but also there are many (1 or 2 out of all) probe sets representing a gene gene behaving discordantly. Should I start looking at the probe set and their match to the desired transcript. Large scale mapping of genes expression values is not allowing me to perform the absolute quantification of genes and the RNA material is from Clinical patients which has been used almost for Microarray hybridizations.

after initial analysis you might wish to analyze the various probe sets for single genes. Look for the annotation, eventually perform a blast to see whether they probe different transcripts. Also note that some probesets simply do not work correctly. This is often difficult to decide but you might be suspicious against low signal probesets if other probesets for the same gene yield a clear signal. On www.genecards.org you'll find confidence values associated with single Affy probesets.

I would suggest to use a custom CDF file instead. This will solve all your questions.

In a few words a couple of ressources propose to use some clean annotation for a lot of microarrays technologies and versions. One of the most used is the BrainArray CDF files (http://brainarray.mbni.med.umich.edu/Brainarray/Database/CustomCDF/genomic_curated_CDF.asp)

or for you

http://brainarray.mbni.med.umich.edu/Brainarray/Database/CustomCDF/14.1.0/entrezg.asp

At the end, for a given gene, all probes (inside a probe-set) were tested, and validated to propose a new (custom) annotation per gene.

This can also easily be used through R and BioConductor

I second Nicolas' suggestion to use the BrainArray CDF. I have gotten much better results over Affy CDF (using the older v2 array).

I believe that v17 of BrainArray is coming out in the next few days.

I agree with the other responses in that you should always use all the probe sets and never average across probes for teh same gene. They may map to the same gene but not the same transcript and may represent either 3' or 5' sequences giving a suggestion of RNA integrity or splice variants. It is much harder to assess individual probes using single arrays and replicates are a better way.

Hi, I think how to conduct the analysis is depending on what is your question and what are you trying to achieve (your ultimate goal). For instance, if you only interested ENSP/ ENST IDs but not the rest, then you can filter the chip based on these IDs (you have to justify why you only interested with Ensembl) and this will leave you a small group of probesets. But regardless what is your question, you should not omit any types of probesets (_at, _s, _a_at, x_at).

When it comes to a more specific question, e.g. multiple ENSP/ENST IDs per probeset, you can double-check if such IDs also appeared on the associated GeneID or gene symbol provided in Affy annotation file using NCBI genbank. By doing this, you can ensure that the provided ENSP/ENST IDs is very likely to be found in such probesets (this will give you confidence on your findings).

I am working on Affy HG-U133_Plus_2 chip and I know it is very difficult to pin down on specicifc IDs (Ensembl, Gene symbol, affy description on cytoband) for probesets as a single probeset happened to be labelled with multiple IDs, or a single IDs is labelled with multiple probesets. Below is my experience on this chip.

My ultimate goal is to identify robust biomarker for cancer diganosis. I started my initial analysis using the entire dataset (54675 probesets). Then, using the result from my initial analysis (the top N statistically significant probesets), I start 'cleaning' my findings, i.e. removing probesets with multiple geneID, identify unknown probesets; using blast search. I perform blast search using both the Public Representative ID and Target probes, verify the blast result using the cytoband info in the affy annotation file. I then filtered my finding based on up-regulation expression on either cancer or healthy groups. Meaning that if a single gene (e.g. A2ML1) in probesets 1553505_at and 1564307_a_at has different expression regulation in cancer group (for ex., 1553505_at has higher fold-change in cancer than in healthy, but 1564307_a_at has lower fold-change in cancer than in healthy), then I will remove this gene from my final results, as this gene does not show persistent up-regulation behaviour on the cancer group.

Hope this will give you some clear idea on your analysis...

Amit,

First look at the Affymetrix Documentation on NetAffx. At the top is a quality score. If the quality is not Aor B, I would ignore it, you can also use this information to see what transcripts hybridize to this probe set. Second, the probe sets for all of the Affymetrix U133 plus 2.0 arrays have been mapped on UCSC genome browser. If you have results that don't agree, do a search for the gene in question and look there for where region that the probe sets cover. To see these probe sets on the genome browser, go down to the Expression section and choose “Pack” for the Affy U133 Plus 2.0 than go to the bottom of the screen and click refresh. Once you turn these on, they will come up every time you use the UCSC browser until you turn it off. In some cases the probe sets is in an intron or outside the gene. If this info is still ambiguous, then go back to NetAffx and look at the specific probes used.

These are all fine suggestions. double check the probe mappings as a few years ago ensemble messed up the probe maps by not taking into account strand, others may have made this mistake as well. Ony thing that was not mentioned is denoising. Typicall half the probe sets have signal that is at or bleow a background level of "expression". Once filtered most genes drop to a single porbe or multiple probe that have a high degree of correlation.

Best

Amit- thanks for adding more context to your question, but I still don't understand why it is so important to resolve these discordant results at a macroscale. Most microarray analyses use some form of clustering to look at gene networks. In this respect it is fine to "let the chips fall where they may."

That is, even if the same gene maps to different clusters due to discordant probe set info, any one gene is given so little weight in the overall analysis, that it is unlikely to throw a wrench in your analysis. Multiple genes in the same network would need to cluster together to hint at a biological insight. Ambiguity about any one member's placement in the cluster should not undermine the whole finding, unless you are working with some sort of hub-gene based analysis.

I also agree with the suggestions that the UCSC and Affy genomic matching have a number of errors and/or omissions. Caviat emptor. But also be careful that you don't spend too much time worrying about a problem that may not impact your final hypotheses or conclusions.

Thanks a lot to all for giving me such wonderful insight about the probe details of Affy genechip. I am trying to consider some of the suggestions which I strongly feel will help me to get out from the problem/confusion. Thanks again for giving your precious time to answer my queries.

I agree with the comments made by Daniel Cohen and others. I would concentrate on finding if there is a considerable variation in gene expression between your 'patients' and 'controls' reproducibly rather than picking a gene or genes without strong justification. I should emphasize choosing your controls appropriately. As indicated by many, there are numerous errors that need to be considered. What is important is the finding of the collection of genes/networks that may be skewed in a certain pathological condition in comparison to actual 'controls'. Again, this depends on the status of the condition-early vs. late stages of the disease under study. There are direct and indirect effects on global gene expression. These microarray studies are also subject to biases and variations at many levels-collection of materials for RNA isolation, RNA storage and handling, technical issues relating to hybridization and importantly at the level of analysis.

As suggested above, treat each one individually. Preprocess, normalize them so that the expression is above a certain threshold as compared to the background. Calculate the differential expression and filter out significant genes with some p-value. Then see which of the probe sets of the same gene appear in the final list and most likely you will get the probe set with matches to the gene.

Dear Aravind, Everything was done in accordingly as mentioned by you which is a normal procedure one should do to get the deferentially regulated probes. I followed some of the suggestions and could able to settle my confusions. Thanks for commenting.

You should use all the probe sets and should not do the average of expression.

Dr. Jagadish, The idea is to map the expression profiles in a gnome wide level to understand the linked processes those are might be getting affected due to P. falciparum infection. I do understand that different probe sets representing a gene might map the alternate transcripts/isoform. Probe sets were mapped to its Ensemble protein ids. To my knowledge ensemble give unique id to each splice variant/isoform. If multiple probe sets mapping to the same ensemble id shows different expression pattern then how to tackle it? When I looked into this, I could find that only three genes (ensemble ids) with multiple probe sets showed the difference where one representing probe set was up regulated and the other one was down regulated and all are A grade probe sets. When I looked into probe sets annotation, I could see that one is specific to the desire transcript (_at) and the other one is although specific but also have the cross hybridization potential with other transcripts derived from different gene/s (_x_at). This allowed me to consider only specific probe sets and overlook the probe sets with cross hybridizing potential which allowed me to settle my earlier confusion. Hope I am following the right path.

Thanks a lot Dr. Zimmermann. Seems like this will allow me to investigate more into probe sets and their performance across a wide category of samples. Hope this will give more confidence in choosing the best probe set for my further data analysis. Thanks again.

I have the same problem in predicting functional associations of other genes with one specific gene. For example, two probesets for the specific gene will results two lists. Which is better, pool the two list together or just the use the overlap of the two?